Handling Multiple Outcomes in Health Data
Sivakamakshi Muthu Kumarasamy, Shivadarshini Sreekanth1
1 School of Mathematical and Statistical Sciences, University of Galway
Background and Problem
Health research increasingly needs integrated analysis of multiple outcomes (e.g., survival, progression, biomarkers, quality-of-life). Traditional approaches fit separate models for each endpoint, implicitly treating outcomes as independent and potentially missing biological and temporal links. Composite endpoints can improve simplicity but may hide heterogeneity across outcome types.
Why Multiple Outcomes Matter
- Clinical trials commonly generate linked endpoints (OS, PFS, biomarkers).
- Separate analyses can ignore dependency, reduce interpretability, and complicate decisions.
- Goal: motivate and build toward multivariate / joint modelling frameworks aligned to disease pathways.
Objectives of Project
- Compare strategies that treat outcomes independently, via composite endpoints, and via multivariate/joint frameworks (with potential extension to multi-task modelling).
- Replace purely univariate workflows with integrated approaches where appropriate.
- Implement joint modelling concepts linking time-to-event and other outcomes via shared latent disease processes.
Data Sources and Datasets
Dataset Overview (Quick Scan)
- SECOMBIT Trial: Phase II randomised; N ≈ 209 metastatic melanoma; 3 arms.
- Endpoints: Overall Survival (OS) and Total PFS.
- Covariates: sites, ULN_LDH, TMB, JAK, PD (as available).
- MSKCC Prostatectomy Cohort: N ≈ 181; multiple time-to-event outcomes.
For the clinical trial component we begin with SECOMBIT (Ascierto et al., 2023; 2024).
For predictive modelling, we plan to evaluate methods on the MSKCC prostatectomy cohort (Taylor et al., 2010).
Descriptive Statistics of Datasets
Early Results and Descriptive Statistics
Key clinical and biological variables are summarised for each dataset. For the SECOMBIT trial, baseline variables include disease burden measures and biomarkers such as Lactate Dehydrogenase (LDH), Tumour Mutational Burden (TMB), and pathway-related markers. LDH is clinically relevant, as elevated levels have been associated with poorer prognosis and reduced response to anti–PD-1 therapy.
Categorical variables are presented using clinically meaningful thresholds (e.g., LDH classified using the Upper Limit of Normal, ULN), while continuous variables are summarised using appropriate descriptive statistics.
Baseline characteristics are summarised to assess group comparability. For SECOMBIT, variables are presented across the three treatment arms, whereas for the MSKCC prostatectomy cohort, summaries are provided for the full cohort (n = 181).
Preliminary time-to-event analyses are conducted for Overall Survival (OS) and Progression-Free Survival (PFS). Survival distributions are visualised using Kaplan–Meier curves, with differences between treatment arms assessed using log-rank tests. Cox proportional hazards models are fitted to estimate covariate-adjusted effects, with model assumptions.
Results are interpreted in both clinical and methodological contexts, highlighting the limitations of analysing multiple outcomes independently and motivating integrated modelling approaches.
Baseline Summary Table (by Arm)
|
A (N=69) |
B (N=71) |
C (N=69) |
Overall (N=209) |
|
|---|---|---|---|---|
| sites | ||||
| 0 (0%) | 1 (1.4%) | 1 (1.4%) | 2 (1.0%) | |
=3 |
26 (37.7%) | 29 (40.8%) | 26 (37.7%) | 81 (38.8%) |
| 1-2 | 43 (62.3%) | 41 (57.7%) | 42 (60.9%) | 126 (60.3%) |
| ULN_LDH | ||||
| 2 (2.9%) | 2 (2.8%) | 1 (1.4%) | 5 (2.4%) | |
| elevated | 26 (37.7%) | 28 (39.4%) | 20 (29.0%) | 74 (35.4%) |
| normal | 41 (59.4%) | 41 (57.7%) | 48 (69.6%) | 130 (62.2%) |
| TMB | ||||
| 41 (59.4%) | 46 (64.8%) | 39 (56.5%) | 126 (60.3%) | |
| <10 | 20 (29.0%) | 17 (23.9%) | 18 (26.1%) | 55 (26.3%) |
=10 |
8 (11.6%) | 8 (11.3%) | 12 (17.4%) | 28 (13.4%) |
| JAK | ||||
| 40 (58.0%) | 47 (66.2%) | 39 (56.5%) | 126 (60.3%) | |
| mut | 5 (7.2%) | 7 (9.9%) | 14 (20.3%) | 26 (12.4%) |
| wt | 24 (34.8%) | 17 (23.9%) | 16 (23.2%) | 57 (27.3%) |
Kaplan - Meier Curve
Figure 1: Overall Survival (OS) by treatment arm.
Figure 2: Progression-Free Survival (PFS) by treatment arm.
Cox model: ARM + sites
Overall Survival (OS)
| Term | HR | CI low | CI high | p |
|---|---|---|---|---|
| ARMB | 0.632 | 0.373 | 1.071 | 0.0882 |
| ARMC | 0.722 | 0.437 | 1.192 | 0.2030 |
| sites>=3 | 1.641 | 1.069 | 2.519 | 0.0235 |
| sites1-2 | NA | NA | NA | NA |
Progression-Free Survival (PFS)
| Term | HR | CI low | CI high | p |
|---|---|---|---|---|
| ARMB | 0.611 | 0.360 | 1.036 | 0.0673 |
| ARMC | 0.725 | 0.439 | 1.198 | 0.2090 |
| sites>=3 | 1.634 | 1.064 | 2.508 | 0.0248 |
| sites1-2 | NA | NA | NA | NA |
Next Project Steps
- Variable discombobulation
- Expand our minds with explosive machine learning
- Explore multi-task learning approaches for multi-endpoint risk prediction
GitHub
The code and datasets for this project can be viewed at our GitHub repository here: https://github.com/darshu-d/MSc-Research-project-
References
Ascierto PA et al., SECOMBIT 4-year results. Nat Commun (2024).
Ascierto PA et al., SECOMBIT sequencing trial. J Clin Oncol (2023).
Taylor BS et al., Integrative genomic profiling of prostate cancer. Cancer Cell (2010).